Classification: NONAUTHORITY / NONINSTALLED CANDIDATE

Target task: BRAIN-CONTROL-PLANE-FAILURE-CLUSTER-REMEDIATION-20260906

This delta does not change canonical Brain authority. It is the protocol half of the tested worker candidate bridge_cp_v8m3.ts.

For op="brain", the Val dispatcher MUST reject jobs whose argument vector cannot be accepted by the current canonical Brain CLI grammar.

Important current examples:

• status takes no extra arguments.
• verify-chain, verify-claim-graph, and recover-stale take no extra arguments.
• renew-lease uses --seconds, not --lease-seconds.
• ensure-task requires --workstream, --dedupe-key, --success, and --value.
• unknown options, duplicate options, missing option values, malformed typed values, malformed JSON-valued options, and missing required options fail before the persistent Brain worker invokes brain.py.

This is transport validation only. It MUST NOT broaden the canonical command surface.

The normal CHAT boot sequence is:

1. submit register-agent;
2. wait until that exact bridge job is COMPLETE with a successful result;
3. only then submit claim --agent <id> --mode INTERACTIVE --lease-seconds <seconds>.

Do not enqueue register and claim as same-second independent READY jobs and rely on created_at,id ordering. Current scheduler ordering can select a lexicographically earlier claim before registration when timestamps and priority tie.

A failed, unknown, expired, or non-COMPLETE registration is not permission to submit the dependent claim.

A completed bridge job is not automatically usable evidence.

For status:

• if outer result_json is invalid: fail closed;
• if stdout_truncated=true: classify STATUS_STDOUT_TRUNCATED;
• if stdout is absent: fail closed;
• if stdout is not complete valid JSON: classify STATUS_STDOUT_INVALID_JSON.

Never parse a 512,000-byte prefix as though it were a complete canonical status document.

A Val durability snapshot is DURABILITY_ONLY_NONAUTHORITY.

It may be used as a compact current-state view only when all of the following hold:

1. fresh canonical verify-chain returns ok=true;
2. computed_head == stored_head;
3. fresh canonical verify-claim-graph returns ok=true and zero problems;
4. the selected snapshot independently verifies its event chain and its content-addressed identity;
5. snapshot event_head_hash equals the canonical verified chain head exactly.

If the heads differ, classify:

DURABILITY_LAG_FAIL_CLOSED

and do not use snapshot tasks, agents, workstreams, generations, or verified-state fields as current canonical state.

Reference candidate: brain_compact_boot_state_v1_v8m3.ts.

At the V8M3 test point:

• canonical verified chain: 1,656 events;
• canonical head: 51fd6222176ae1fcff06510ec3a688b1140f5ef3dcbec55a073529d5bf835991;
• claim graph: 8 claims, zero problems;
• latest verified snapshot: generation 1502 / 1,548 events;
• snapshot head: ef6ffd252b0e73a9240647dab0630503841074c862c6346806606b494ea25dd3;
• therefore compact state classification: DURABILITY_LAG_FAIL_CLOSED;
• latest full status output is transport-truncated at 512,000 stdout bytes.

Do not install this delta until:

• the canonical task is actually leased to the implementing agent;
• exact candidate hashes are preserved;
• bridge regression remains green;
• production worker/protocol diff is bounded to the proven repair;
• fresh chain and claim-graph checks pass after installation;
• no direct Sprites hot-path dependency is introduced;
• no human_authority_approval capability is claimed.

Current tested identities:

• guard: dc5594135239f9aac6dfc639b730c1e5c91f4e28e51637ee746098d2f28f0323
• compact boot: 71fe6fd715f781dcdcc679264386eb8e18ae2dfc46c9c153fe032c3a7e5dc307
• worker candidate: c6adecc717fe17b8c94180e59cb5bd4f94e928aa0e73c7df79b45f9204360dea
• worker regression: 52/52 PASS
• two-sided queue replay: 800/800 recent successful supported jobs accepted; 18/18 genuine CLI/dispatch failures rejected pre-host
• replay script: 6d1daa2f5c6d8300afd8677351c204d9f313010dc5e975f4218ce1c9509ba6f0

Historical note: guard v1 d16b944491880c0c6eeb1dc7c6d10ba11333d3d4d2dbfff42e1aba495b6180df was falsified for one compatibility case: canonical ensure-task --help had previously completed successfully while v1 rejected it. V2 preserves exclusive --help compatibility and passes the two-sided replay.